Download presentation
Presentation is loading. Please wait.
Published byOddbjørn Arntsen Modified over 5 years ago
1
Cascade R-CNN: Delving into High Quality Object Detection
Zhaowei Cai, Nuno Vasconcelos University of California San Diego
2
Object Detection Definition of positive/negative samples positive??
Ambiguous!! positive??
3
Object Detection Standard definition
is to use an intersection over union (IoU) threshold (usually 0.5). positive √ positive √ positive × positive ×
4
Object Detection During training IoU=0.5
these samples are on opposite sides of the class boundary. IoU=0.5
5
Object Detection However, this threshold is loose IoU=0.5
many positive samples are of low quality. IoU=0.5 low quality positives
6
Current Object Detection
Detection results can be quite noisy! low quality detector
7
High Quality Object Detection
Q: how to train a high quality object detector? high quality detector
8
Low Quality Object Detection
Training with low IoU threshold IoU=0.5
9
High Quality Object Detection
Training with high IoU threshold IoU=0.7
10
High Quality Object Detection
Two problems 1. Training overfitting, due to exponentially vanishing positive samples. 2. Inference-time quality mismatch between the detector and its input proposals.
11
Cascade R-CNN Multi-stage framework, increasing IoU threshold
Faster R-CNN Cascade R-CNN
12
Why Cascade R-CNN? Two benefits
IoU=0.5 IoU=0.6 IoU=0.7 Two benefits 1. Reduce training overfitting, since the proposal quality is sequentially improved. 2. Reduce Inference-time quality mismatch, because the same cascade is also used at inference.
13
Effective High Quality Detection
Large gains for high quality detection Detection Quality
14
State-of-the-art Results
The best single model results on COCO
15
Cascade R-CNN: Delving into High Quality Object Detection
Welcome to our poster! (P2-3-D14) Cascade R-CNN: Delving into High Quality Object Detection
Similar presentations
© 2025 SlidePlayer.com Inc.
All rights reserved.